Open-Set Plankton Recognition

Joona Kareinen, Annaliina Skyttä, Tuomas Eerola, Kaisa Kraft, Lasse Lensu, Sanna Suikkanen, Maiju Lehtiniemi, Heikki Kälviäinen · 2025

Read Open-Set Plankton Recognition by Joona Kareinen, Annaliina Skyttä, Tuomas Eerola, Kaisa Kraft, Lasse Lensu, Sanna Suikkanen, Maiju Lehtiniemi, Heikki Kälviäinen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-11318.txt.

Introduction

Joona Kareinen and co-authors work on naming plankton from photos when new kinds show up, in a 2025 paper. Plankton are tiny water life that feed sea food webs and track ocean health. Cameras now gather huge photo sets, but they also catch bits that are not plankton and kinds not seen in training.

The team tests three ways to spot and set aside unknown kinds while still naming known ones well. They use both plant and animal plankton photos and vary the line at which a photo is set aside. High scores were reached with the best settings. The data set was shared in public so others can build on the work.

Catalog record

CREATORJoona Kareinen, Annaliina Skyttä, Tuomas Eerola, Kaisa Kraft, Lasse Lensu, Sanna Suikkanen, Maiju Lehtiniemi, Heikki Kälviäinen
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-11318

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Open-Set Plankton Recognition Joona Kareinen, Annaliina Skyttä, Tuomas Eerola, Kaisa Kraft, Lasse Lensu, Sanna Suikkanen, Maiju Lehtiniemi, Heikki Kälviäinen Abstract This paper considers open-set recognition (OSR) of plankton images. Plankton include a diverse range of microscopic aquatic organisms that have an important role in marine ecosystems as primary producers and as a base of food webs. Given their sensitivity to environmental changes, fluctuations in plankton populations offer valuable information about oceans’ health and climate change motivating their monitoring. Modern automatic plankton imaging devices enable the collection of large-scale plankton image datasets, facilitating species-level analysis. Plankton species recognition can be seen as an image classification task and is typically solved using deep learning-based image recognition models. However, data collection in real aquatic environments results in imaging devices capturing a variety of non-plankton particles and plankton species not present in the training set. This creates a challenging fine-grained OSR problem, characterized by subtle differences between taxonomically close plankton species. We address this challenge by conducting extensive experiments on three OSR approaches using both phyto- and zooplankton images analyzing also on the effect of the rejection thresholds for OSR. The results demonstrate that high OSR accuracy can be obtained promoting the use of these methods in operational plankton research. We have made the data publicly available to the research community. 1 Introduction Most image recognition models are trained based on the closed-set assumption where the test set is expected to contain the same set of classes as the data the model was trained on. However, this assumption does not hold in most real-world applications where the recognition model can be expected to encounter images from various previously unseen categories. This leads to an open-set scenario where the recognition model should be able to not only recognize the known classes but also identify if an image is from a previously unseen class. A good example of this is plankton species recognition from images. Automatic plankton imaging instruments capture various non-plankton particles when deployed in real aquatic environments. Moreover, plankton species vary between geographic regions, leading to datasets with only partly uniform class composition. To advance the use of plankton recognition models operationally, the open-set recognition (OSR) problem should be addressed. Plankton are microscopic organisms that live in aquatic environments, drifting and swimming along tides and currents. Plankton consists of two main groups: phytoplankton and zooplankton [ 8 ] . Phytoplankton, which obtains energy through photosynthesis, produces oxygen and organic compounds, which serve as an essential food source for various microorganisms. On the other hand, zooplankton are a primary food source for many animals and have a key role in nutrient cycling by consuming phytoplankton and transferring energy to higher trophic levels, such as commercially important fish species [ 10 ] . Plankton is vital for the marine food web as primary producers and indicators of ocean health, responding quickly to changes in temperature and currents [ 7 , 11 ] . Monitoring how plankton populations change can provide valuable information about the effects of climate change [ 11 ] , driving the increasing collection of information about plankton concentrations and species distribution. Recent technological advancements have led to automatic and semi-automatic plankton imaging systems, used to gather massive plankton datasets [ 17 ] . Going manually through these images is infeasible, and to fully utilize the available data, an automatic system recognizing plankton species with minimal supervision is required. Several challenges exist in plankton recognition including subtle visual differences between species, varying image quality, unrecognizable particles, and class imbalance [ 6 ] . A large number of plankton recognition methods are available [ 6 ] . These vary from traditional feature engineering-based approaches [ 31 , 19 ] to convolutional neural networks [ 18 , 4 , 14 ] , and lately, vision transformers [ 16 ] . Despite the high recognition accuracies reported in studies, these methods have not been widely adopted for operational use. One notable reason for this is the open-set nature of the recognition task, which is not properly considered in most studies. Works that do propose a method for open-set plankton recognition [ 1 , 27 ] report notable drops in recognition accuracy compared to closed-set recognition models. Plankton image data introduces various challenges, such as the fine-grained nature of the recognition task and label uncertainty. These factors render the open-set plankton recognition problem more difficult than many other domains in which OSR methods have been traditionally developed. However, this also makes plankton recognition an interesting testing ground for developing general-purpose OSR methods. In this paper, we conduct extensive experiments on three different OSR approaches for both phyto- and zooplankton images. The methods selected for the study are: (1) OpenMax [ 3 ] allowing the identification and handling of unknown classes by modifying the output probabilities, and two metric learning approaches: (2) ArcFace [ 5 ] and (3) Class Anchored Clustering (CAC) [ 22 ] . Metric learning (see Fig. 1 ) allows mapping images to a metric space where images from the same class are close to each other, enabling the recognition and identification of unknown classes simultaneously by setting distance (rejection) thresholds. For ArcFace, we utilize the implementation from [ 1 ] , developed for open-set plankton recognition. CAC has not been previously applied to plankton recognition. Figure 1 : Open-set plankton recognition using metric learning. Classes 1–4 represent the known classes, and ? unknown classes. Due to the subtle differences between plankton species, the rejection thresholds have a crucial role in OSR as they form the basis for identifying unknown classes. We consider two approaches for selecting the rejection threshold values: (1) setting a fixed threshold for all classes based on validation set, and (2) class-specific thresholds. To obtain class-specific threshold values, we utilize a quantile-based approach that does not require examples from unknown classes and can be used to find thresholds for all the classes based on a single quantile value minimizing the number of method hyperparameters. Furthermore, we study the sensitivity of the method to the threshold values. In the experimental part, we show that while the class-specific threshold values increase the accuracy of the methods on average, high accuracy can be obtained with a fixed threshold value when selected properly. The main contributions of this study are as follows: (1) a new publicly available SYKE-plankton_ ZooScan_2024 zooplankton dataset [ 12 ] that together with the existing SYKE-plankton_ IFCB_2022 phytoplankton dataset [ 15 ] provides a versatile and challenging dataset to develop and test plankton recognition and OSR methods, (2) the first application of CAC for open-set plankton recognition, and (3) extensive experiments on three OSR methods on plankton data including analysis of the effect of rejection thresholds. 2 Related work 2.1 Plankton recognition A large number of automatic plankton recognition methods have been proposed to reduce manual work and to analyze the vast amount of plankton data. Early plankton recognition methods relied on classification techniques using handcrafted features. These feature engineering methods often combined various types of features ( e.g . , shape, texture, and local features) to obtain more representative feature vectors, and traditional classifiers such as support vector machines (SVMs) or decision trees to predict the class labels. For example, one

Read the complete text